RankNet
Asking a model how much a user will like an item is a harder question than asking which of two items they will prefer — and the second question is the one a ranked list actually needs answered. RankNet was the model that took that shift seriously.
Learning from pairs
RankNet is a neural network that scores one item at a time but trains on pairs. For two candidates with a known preference — one clicked, one skipped — it turns the difference of their scores into a probability with a sigmoid, then minimises cross-entropy against the observed preference. Absolute score values are left unconstrained; only the gap between them carries meaning, which is exactly the right contract for a list that will be sorted and truncated.
Why it mattered here
- Implicit feedback offers no ratings to regress on, but it hands out pairs for free: a watched title against one shown and ignored.
- Comparisons are drawn within a single user's slate, so a user who clicks on everything and one who clicks rarely contribute equally.
- Training is ordinary backpropagation, so the scorer can be any architecture — including the MLP sitting on top of a ranking stage.
Its limit, and its heirs
RankNet treats every pair as equally important, so fixing a swap at position 200 earns the same reward as fixing one at position 2. LambdaRank patched this by weighting each pair's gradient by the NDCG change the swap would cause, and LambdaMART carried the idea into gradient-boosted trees. RankNet survives inside both as the pairwise loss, and as a default objective in ranking libraries such as TensorFlow Ranking.
RankNet is a neural ranking model trained on pairs of items rather than on individual targets: the difference between two items' predicted scores is passed through a sigmoid and fitted by cross-entropy to the observed preference. It made implicit feedback directly trainable for ranking, since skipped-versus-clicked pairs are abundant where ratings are not, and its pairwise loss still sits underneath LambdaRank and LambdaMART in the ranking stage.
Ask someone to score a film out of ten and they will hesitate. Ask which of two films they would rather watch tonight and the answer comes instantly.
RankNet is built on that second, easier question. Instead of guessing a rating for every item, it learns from comparisons: you watched this one and scrolled past that one, so this one belongs higher. Millions of those small comparisons add up to a sense of order, and an order is all a list of suggestions really is. It is one of the older ideas in this area, and most of today's ranking systems are descendants of it.